Exploring Abstract Pattern Representation in The Brain and Non-symbolic Neural Networks

Area of Science:

• Cognitive Science
• Neuroscience
• Artificial Intelligence

Background:

• Human cognitive and linguistic generativity relies on identifying abstract relationships.
• Previous research indicated simple recurrent neural networks (SRNs) struggled with abstract pattern generalization, unlike human infants.

Approach:

• Developed deep Long Short-Term Memory (LSTM) models trained on cochleagrams representing auditory stimuli.
• Investigated models trained to identify syllable trigram words and reduplication patterns.
• Utilized representational similarity analysis (RSA) to compare model representations with brain activity.

Key Points:

• LSTM models successfully identified abstract syllable repetition patterns and words.
• Discovered representations supporting classification of abstract repetition patterns.
• Found significant correlations between model representations and brain activity in posterior temporal regions.

Conclusions:

• Associative mechanisms in deep learning models can account for abstract pattern recognition.
• These findings challenge the notion that purely associative models are inadequate for explaining cognitive generativity.
• Suggests discoverable representations capturing abstract properties are key to cognitive generativity.